Atlas

Benchmarks

← All benchmarks

TextVQA (validation)

Multimodal · 2019-04-18

TextVQA asks questions about natural images that can only be answered by reading the text in the scene. This row is the 5,000-question validation split, the split model reports conventionally use, scored by the standard soft VQA accuracy against ten human answers (full credit when at least three annotators gave the answer). Higher is better.

Top models (higher is better)

ModelScore
Gemini 3.5 Flash-Lite88.9
Gemma 4 12B IT88.4
Qwen3.5-4B88.0
LFM3-VL-2B87.9
GPT-5.6 Luna85.5
Qwen2 VL 72B Instruct85.5
Nemotron Nano 12B V2 VL BF1685.4
Qwen3.5-2B85.0
Qwen2.5-VL-7B-Instruct84.9
LFM2.5-VL-3B84.3
Qwen2 VL 7B Instruct84.3
Qwen2.5-VL-72B-Instruct83.5
InternVL2.5-78B83.4
Qwen3-VL-8B-Instruct82.9
Gemma 4 E2B IT80.6
MiniCPM-o 2.680.1
MiniCPM-V 2.680.1
Qwen2 VL 2B Instruct79.7
Qwen VL Max (rolling alias)79.5
Qwen2.5-VL-3B-Instruct79.3
Gemini 1.5 Flash (May 2024 report)78.7
Gemini 1.5 Pro (May 2024)78.7
Gemini 1.5 Pro (February 2024 report)73.5
Gemini Nano 265.9
Gemini Nano 162.5
Loading Atlas data…